Neuro-fuzzy model based on digital images for the monitoring of coffee bean color during roasting in a spouted bed

• Published in: Expert systems with applications Vol. 54; pp. 162 - 169
• Main Authors: Virgen-Navarro, Luis, Herrera-López, Enrique J., Corona-González, Rosa I., Arriola-Guevara, Enrique, Guatemala-Morales, Guadalupe M.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.07.2016
• Subjects: Artificial neural networks; Beans; Coffee; Coffee roasting; Color; Digital image; Fuzzy logic; Fuzzy set theory; Mathematical models; Moisture content; Monitoring; Neuro-fuzzy model; Roasting; Spouted bed; Coffee roasting; Neuro-fuzzy model; Digital image; Spouted bed; Moisture content; Color
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2016.01.027

•Coffee color during roasting in a spouted bed was assessed by an ANFIS model.•The input variables were color parameters obtained from digital images.•For monitoring in the process line, we constructed an image-capture device.•The model's goodness of fit was assessed by R2, RMSE and SRIC.•The model is proper for monitoring in the process line. An adaptive-network-based fuzzy inference system based on color image analysis was used to estimate coffee bean moisture content during roasting in a spouted bed. The neuro-fuzzy model described the grain moisture changes as a function of brightness (L*), browning index (BI) and the distance to a defined standard (ΔE). An image-capture device was designed to monitor color variations in the L*a*b* space for high temperatures samples taken from the reactor. The proposed model was composed of three Gaussian-type fuzzy sets based on the scatter partition method. The neuro-fuzzy model was trained with the Back-propagation algorithm using experimental measurements at three air temperature levels (400, 450 and 500°C). The performance of the neuro-fuzzy model resulted better compared to conventional methods obtaining a coefficient of determination > 0.98, a root mean square error <0.002 and a modified Schwarz–Rissanen information criterion <0. The simplicity of the model and its robustness against changes in the input variables make it suitable for monitoring on-line the roasting process.